TLDR: A new study developed an AI model to predict round outcomes in VALORANT by analyzing minimap information from match footage. By incorporating detailed tactical features like enemy detection and skill usage, the model achieved approximately 81% prediction accuracy, significantly outperforming models that only used raw minimap data, especially in the middle and late phases of a round. This highlights the importance of visual tactical information for understanding complex esports strategies.
Esports has grown into a global competitive arena, captivating large audiences, especially in genres like First-Person Shooters (FPS) such as VALORANT. These games demand intricate strategies, seamless teamwork, and precise individual skills. However, for many viewers, particularly those new to the scene, understanding the subtle shifts in a match and how player decisions lead to wins or losses can be challenging. This complexity can hinder the further expansion of the esports fan base.
To enhance the audience experience, real-time match outcome prediction has emerged as a valuable tool. While many existing studies in games like Multiplayer Online Battle Arenas (MOBAs) rely on match log data and statistical information, these approaches often miss the rich tactical details visible on screen, especially on the minimap. In FPS games like VALORANT, visual cues such as the positions of teammates and opponents on the minimap are crucial and significantly influence the outcome of a round.
A New Approach to VALORANT Round Prediction
A recent study proposes an innovative method for predicting round outcomes in VALORANT by directly analyzing match footage, with a particular emphasis on visual tactical features extracted from the minimap using deep learning. The researchers adopted TimeSformer, a Transformer-based architecture known for its strength in understanding spatio-temporal contexts, as their foundational model.
The core idea is that explicitly modeling tactical features will lead to more accurate predictions. To test this, two variants of their model were designed and trained on datasets derived from official VALORANT tournament broadcasts.
Dataset and Methodology
The study collected 1,376 broadcast videos from 2024 VALORANT tournaments, encompassing 29,506 rounds. After filtering for rounds within the standard 100-second duration, 21,229 rounds were used. Notably, tournament footage provides a unique advantage by displaying the positions and skill information for both teams on the minimap, unlike normal match footage which only shows the player’s team.
Two datasets were constructed:
- Dataset A: Minimap Information – This dataset contained raw minimap information, such as character positions, segmented into individual rounds and labeled with win or loss outcomes and the map played.
- Dataset B: Minimap Information Augmented with Tactical Events – Building on Dataset A, this enhanced dataset included additional labels for tactical events. These events signify moments when teams gain or lose tactical information, often inferred from visual or auditory cues within the game. Examples include detecting enemy presence through footsteps (using their fixed audible range) and monitoring skill usage to label moments of intelligence gained or denied.
The models, implemented in PyTorch, used a TimeSformer architecture with default settings. Input videos were downsampled, and tactical event features in Dataset B were aggregated and fused with visual features before being fed into the TimeSformer model.
Key Findings and Improved Accuracy
The evaluation compared Model A (trained on Dataset A) and Model B (trained on Dataset B) based on their per-second accuracy of round outcome predictions. The results were compelling:
- Model B, which incorporated additional tactical event labels, achieved an overall prediction accuracy of approximately 80.55% across the entire round.
- This was a significant improvement of about 8 percentage points over Model A, which only used raw minimap information, achieving 72.28% accuracy.
Furthermore, the analysis showed that Model B consistently outperformed Model A from approximately 24 seconds after the start of the round. From 30 seconds into each round, Model B’s accuracy soared to over 80%. This is particularly noteworthy because in VALORANT, key abilities that heavily influence a round’s trajectory are often activated during the middle phases. The findings suggest that the additional temporal and spatial information provided by tactical events plays a crucial role in interpreting game states, especially after abilities are deployed.
Also Read:
- Enhancing AI Strategy in Complex Games by Retaining Historical Information
- Unpacking AI’s Intent: A User Study on Explaining Reinforcement Learning Goals
Implications and Future Directions
The study highlights that detailed tactical event labels, including spatial information and inferred skill usage, are critical features for accurately predicting round outcomes in VALORANT. These insights, extracted directly from match footage, capture nuances that simple minimap information alone cannot.
While the method has achieved a commendable level of accuracy, there is room for further refinement. Future work aims to enrich input features with elements like movement vectors and player-specific historical data, and to refine the model architecture. Ultimately, the researchers envision extending this system to provide coaching for non-professional players, offering analysis and explanations of effective strategies and decision-making processes observed in professional matches, thereby fostering a deeper in-game strategic understanding.
For more detailed information, you can read the full research paper here.
